Everything your organisation knows, in one question. And only what the asker is allowed to see.
Ask Todo answers questions about the work itself — what was decided, why something is stuck, what changed while you were away — with a link to the source behind every claim. It lives inside the system the work already runs in, and enforces the permissions you already set.
The launch was pushed back over the image-loading bug opened on 12/03, still open1
The call summary from 18/03 agreed to freeze the gallery until it closes2
A knowledge assistant that leaks once is finished.
It is the number-one objection in this category, and rightly so: the moment an assistant hands someone a document they were never meant to see, no apology undoes it. So permissions here are not a feature, they are the structure — and we would rather explain how it works than promise you it's safe.
The knowledge exists. It just isn't reachable by anyone who needs it.
The decision is in a comment from eight months ago, the agreement is in a call transcript, and the reasoning is in a file somebody once uploaded. The person who knows where — is on holiday.
From a plain-language question to an answer you can check
Not chat-with-your-documents — a knowledge layer over the work
The sources are what already happens in your system: tickets and their comments, chat conversations, call transcripts and the files attached to them.
Ask in plain language
"What was decided about pricing", "why is the launch slipping", "what moved this week" — ordinary phrasing, no search syntax and no keywords.
A clickable citation per claim
Clicking a source opens the item itself inside Todo — the ticket, the message or the file — not a text panel you cannot verify.
"Summarise" without phrasing a question
One button on a long ticket or a conversation. It is the button people actually press, because it doesn't require knowing what to ask.
Numbers from the data, not the model
"What's overdue", "who owns this", "how much time went in" are answered by querying the data — a model shouldn't guess a number you can count.
"What changed while I was away"
Every item carries its date, dates render in your timezone, and "today" is passed to the model explicitly instead of being guessed.
From a spec document to a task list
Point at a document or a thread and get proposed tasks with titles and descriptions — for you to approve before any are created.
A private space only you can reach
Notes and files you put there yourself, spanning projects. Nobody else is ever retrieved into them — not a manager, not an administrator.
An admin who decides what goes in
On and off, which sources get indexed, whole projects excluded, and a coverage report showing what was taken in, what failed and why.
Most search engines simply cannot see half the words in Hebrew
A database's standard full-text search ignores short words — in Hebrew that quietly drops some of the most common ones — and it has no concept of prefixes, so "in-the-project" and "project" are unrelated words to it. In Hebrew that is not a small gap; it is half the questions.
One system, four entirely different questions
A scope is not four products but four filters over the same index — and each one only narrows. No scope ever widens what you are allowed to see.
"I didn't find it" is a valid answer, not a failure
An assistant that invents costs more than one that occasionally admits it has nothing. So the refusal here is a designed, deliberate output rather than an edge case somebody forgot.
Knowledge is only worth something once it moves
This is the difference between enterprise search that finds a document and an assistant that sits inside the work: the answer arrives somewhere you can immediately do something with it.
You decide what goes in — and you see exactly what was spent
The two questions that stall projects like this are "what did it actually read" and "what will this cost me this month". Both are answered on a screen, not in a call with support.
Organisations where the answer exists — with somebody else
The more people, projects and history you have, the wider the gap between what the organisation knows and what an employee can find.
Everything you wanted to know
No, and we would rather say so up front. Ask Todo answers on the work already inside Todo: tickets and their comments, chat conversations, call transcripts and the files attached to them. It is not a broad enterprise search tool indexing every system you own. The practical distinction: broad search is good at discovering scattered information, and an assistant living inside the work system is good at understanding context and carrying on into action. If most of your work runs in Todo, this is exactly its case.
Two ways that reinforce each other. First: permission is computed at question time against the source tables rather than stored ahead of time on the indexed content — so removing someone from a project takes effect on their very next question, with no staleness window. Second: the index stores no field that participates in a permission decision, precisely so none can go stale and widen access by accident. Before every release, 17 tests each try to leak something specific; if one fails, nothing ships.
No. It is the one scope in the system with no back door — not for a manager, not for an administrator, and not through another user's export request. You can see, edit and delete every item in it. Just as importantly, the personal space indexes nothing on its own: only what you explicitly put there goes in.
Really in Hebrew, and it was a separate piece of work. A database's standard lexical search drops short words from the index and cannot recognise prefixes, so the prefixed and unprefixed forms of the same Hebrew word are strangers to each other. We built our own tokenizer that strips prefixes and their chains, normalises final-form letters and removes vowel points and quote marks, and ranks with BM25 — alongside a glossary you define for the acronyms only your organisation understands.
Several things together. Every factual claim must carry a citation, and a sentence without a source simply isn't written. Questions with an exact answer in the data — what's overdue, who owns it, how much time was logged — are answered by querying the data rather than the model, because a number you can count shouldn't be guessed. And when there isn't enough material, the answer is "I didn't find that in what I have access to", together with what was found. The refusal is designed, not a failure.
No, and that is a stated red line rather than an omission. There are no productivity scores, no employee rankings and no "who works slowly". Even in personal memory, a fact of the form "usually runs late" is forbidden — that is monitoring, not memory. The product answers questions about the work; it does not offer opinions about the people.
The model is credits, and the rule is simple: one question is one unit, and that ratio is pinned rather than floating against our costs. The screen shows how many questions remain, when the allowance resets, a 30-day breakdown by user and by project, and a full transaction log — so "where did the credits go" is answered on a screen. Prices and packs are settings that get updated from time to time, so we won't print a number here that goes stale — we'll walk through them together on the demo.
Ask it about your real project, and see what comes back
Leave your details and we'll switch it on for one scope of yours, show how permissions behave across two different users, and answer questions you choose.